Researchers have released SULAND v2, an improved dataset for training object detection models to identify surface landmines from RGB imagery captured by unmanned aerial and ground vehicles. The original SULAND dataset contained numerous annotation errors, including missing or false annotations, localization inaccuracies, and inconsistent class labeling. SULAND v2 addresses these issues by manually re-annotating 33,771 images with 12,433 bounding boxes, ensuring greater accuracy and consistency. Benchmarking 35 detector configurations on this refined dataset showed significant improvements in performance, with YOLOv12-Small achieving the highest in-distribution accuracy and RF-DETR-Large excelling in out-of-distribution scenarios, highlighting that high in-distribution accuracy does not guarantee real-world operational readiness. AI
IMPACT Improves the reliability of AI models for critical tasks like landmine detection, potentially enhancing safety and operational effectiveness.
RANK_REASON Publication of a refined dataset and benchmark for object detection in a specialized domain.
- PFM-1
- PMA-2
- SULAND v2
- unmanned aerial vehicle
- unmanned ground vehicle
- YOLOv12-Small
- YOLOv8
- Label Studio
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
